Upload
others
View
8
Download
0
Embed Size (px)
Citation preview
THE AUTOX
PUBLISHED 2018
SAFETY FACTOR
Democratizing Autonomy
Bringing Self-driving Technologies to Everyday Life
The AutoX Safety Factor 3
Cars play a central role in our lives. Each year, we spend an increasing
number of hours behind the wheel, a trend that may continue unless
drastic changes are made. Yet when we consider how much time is
spent driving to meet our needs, an unmistakable pattern emerges:
driving is rarely more than just a means to an end.
In the world that AutoX envisions, autonomous vehicles connect us
by allowing us to transport physical items to and from one place to
another. Even today, lending a friend a salad bowl, a socket wrench,
or an Xbox controller still means a visit to the purgatory of traffic. By
handling the less pleasant part of the process, AutoX’s autonomous
vehicles can make the thought of sending a gift basket to your friend
the warm, happy one that it always intended to be.
Communities are strengthened by passionate people who strive
to share their works of love with others. The local artisanal baker
who has been exploring new recipes will be overjoyed to send their
regulars samples via autonomous vehicles as soon as the first batch
is out of the oven. Florists, coffee shops, restaurants, and nearly every
other business will benefit from our ability to share their products
and creations with their communities in a safe, affordable, and pain-
free way - and so will the communities that they serve. With AutoX
providing the means, the ends are truly limitless.
On average, every cubic meter of goods travels 2.1 miles to be delivered in the traditional B2B
model1.
2.1 Miles
In New York City, drivers on average spend 107 hours per year searching
for parking spots2.
107 Hours
The average person spends 60 hours per year shopping in grocery stores3.
60 Hours
SERVING OUR COMMUNITY WITH
AUTONOMOUS DELIVERY
AutoX envisions a world where self-driving cars take care of the means, freeing you up to enjoy the ends.
The AutoX Safety Factor 4
Not long after the first motorcar was invented, human
beings began to consider the possibility of making them
autonomous. In 1925, inventor Francis P. Houdina publicly
demonstrated his radio-controlled driverless car, American
Wonder, in the streets of New York City, traveling up
Broadway and down Fifth Avenue through heavy traffic4.
This marked the beginning of humankind’s aspiration for
driverless vehicles. Since 2013, more and more pioneers
have spearheaded the efforts of making self-driving cars a
reality. Level 2 and Level 3 assistive-driving vehicles5 are now
being driven on the road. It seems inevitable that manual
driving will someday be a thing of the past. Rather than
driving to get the things we need, we have them come to us
autonomously.
The next few decades will represent a transition period
where humans and fully autonomous A.I. drivers are
THE WORLD IN
TRANSITION
First Self-driving Car System on the Road
(Dean Pomerleau and Todd Jochem)
All Human Drivers Human Drivers and A.I. Drivers
All A.I. Drivers
Big Manufacturers Dive In
We are here Fully Autonomous Vehicles and Smart
Traffic Network
1995 2013 2018 Future
driving together, co-existing and sharing the road. The
U.S. Department of Transportation has decreed that
the freedom of the open road will be protected and
enhanced. This includes the freedom for everyone to
drive their own vehicles6. The unprecedented sharing
of our roads between humans and computers is the
reason AutoX is pushing the limits of cutting-edge A.I.
technologies to ensure the safety of everyone involved.
We have developed a smart hybrid system, which
introduces a vital human element into the autonomous
driving technology, so that we may better serve our
community.
Autonomous driving technology, like most life-changing
innovations, unleashes limitless possibilitites for the
future. AutoX is pushing hard to build safe self-driving
cars that will share the road with everyone.
The AutoX Safety Factor 5
We composed this voluntary safety self-assessment report to share our
approach to achieving safe autonomous vehicle testing and future deployment
based on the voluntary guidance of A Vision for Safety, Automated Driving System
2.0 and Preparing for the Future of Transportation, Automated Vehicles 3.0. We
open with a general introduction to AutoX’s approach to building a safe and
reliable self-driving delivery platform. Next, we highlight our redundancy-
based system design for ensuring safety. Then, we introduce our testing and
validation methods as well as our safety process and policy. Finally, we address
considerations in public safety on the road, including our designed post-crash
behavior and our engagement plan for first responders.
THE VOLUNTARY SAFETY SELF-ASSESSMENT REPORT
1. Statista. https://www.statista.com/statistics/781119/parcel-delivery-mileage-urban-commercial-transport/2. Drivers spend an average of 17 hours a year searching for parking spots. https://www.usatoday.com/story/
money/2017/07/12/parking-pain-causes-financial-and-personal-strain/467637001/3. Who Does the Grocery Shopping, and When Do They Do It. Jack Goodman. The Time Use Institute. http://
www.timeuseinstitute.org/Grocery16paper.pdf4. Time Magazine. Science: Radio Auto. Aug 10, 1925. Retrieved 29 September 2013. 5. SAE autonomy scale. Level 2: Driver-assist systems that control both steering and acceleration/deceleration.
These systems shift some of the workload away from the human driver, but still require that person to be attentive at all times. Level 3: Vehicles that can drive themselves in certain situations, such as in traffic on divided highways. When in autonomous mode, human intervention is not needed. But a human driver must be ready to take over when the vehicle encounters a situation that exceeds its limits.
6. U.S. Department of Transportation, Preparing for the Future of Transportation, Automated Vehicles 3.0
The AutoX Safety Factor 6
System Safety
The key to achieving system safety is to establish a robust
design process and follow a strict validation process.
We are setting the highest standards in building our self-
driving cars with design redundancies and safety strategies
to handle autonomous driving system malfunctions and
errors. AutoX’s entire system features these strategies.
In the How Our Self-driving Car “Sees“ section, we share
our methodology for designing and building our sensors
and other equipment that enable our self-driving cars to
“see”. In the Full Stack Redundancy chapter, we discuss our
design principle of providing redundancy for the entire self-
driving system, from hardware to software to operations.
In the Testing and Validation section, we share our data-
driven development methods and our rigorous testing and
validation process.
Operational Design Domain
A closed set of operational criteria and constraints, such
as geographic area, weather condition, time of a day, etc.,
are defined as the Operational Design Domain (ODD). The
ODD is extensively tested so that we are certain that our
autonomous driving system is able to operate safety within
it. Different testing or deployment stages have different
ODDs. In the Operational Design Domain section, we
define the ODD for our self-driving fleet.
Object and Event Detection and Response
Object and event detection and response refers to
the detection by the autonomous driving system of
circumstances that are relevant to the immediate driving
task, as well as the implementation of the appropriate
system response. In the Algorithm Redundancy and A.I.
Redundancy sections, we introduce our unique way of
safely handling and responding to immediate driving tasks.
In the The Brain of Our Cars section, we present selected
highlights of our A.I. technology that build a better and
safer self-driving platform.
Fallback (Minimal Risk Condition)
Fallback is the transition to the minimal risk condition
when a problem is encountered. In the Fallback (Minimal
Risk Condition) section, we discuss our implementation of
the minimal risk condition transition process.
Human Machine Interface
Human Machine Interface (HMI) describes a series of
software and hardware user interfaces that help first
responders, customers, and other road users to obtain
necessary information about our self-driving car, as well
as a way for them to safely engage with it. We address
this topic in depth in the Human Machine Interface (HMI)
section. The HMI also covers the interaction between
the vehicle and the remote operator, and is extensively
covered in the Human Remote Operator section.
Cybersecurity
AutoX has followed a thorough process to minimize the
risks of cybersecurity threats. This is especially important
since AutoX has built a teleoperation system which allows
human operators to control the vehicles remotely. In the
Cybersecurity section, we have introduced our methods
to ensure the cybersecurity of the vehicle and the remote
command center.
In the guidance document Automated Driving Systems 2.0: A Vision for Safety,
the National Highway Traffic Safety Administration (NHTSA) has outlined 12
topics that the self-driving system manufacturer should pay attention to for
achieving safety. Below, we outline the sections in our report in which each
of these topics is addressed.
Safety Elements
The AutoX Safety Factor 7
Crashworthiness
Crashworthiness is the consideration and practice of how
to best protect vehicle occupants in a collision. AutoX
integrates its self-driving technology with Lincoln MKZs and
Chrysler Pacificas, which meet applicable requirements
of the Federal Motor Vehicle Safety Standards (FMVSS)
as issued by the National Highway Traffic Safety
Administration (NHTSA).
Post-Crash ADS Behavior
A well-designed post-crash behavior ensures that the
autonomous driving system returns to a safe state
immediately after being involved in a crash. AutoX has
standardized the behavior of the vehicle after a crash,
which minimizes the risk to first responders and other road
users. In the Post Crash Behavior section, we provide a
brief introduction to the desired post-crash behavior of our
self-driving cars.
Data Recording
Learning from crash data is a central component to the
safety potential of the autonomous driving system. In
the Data Recording section, we present our strategies
to ensure the recording, security, and proper usages of
testing data from crashes.
Customer Education and Training
Education and training are imperative for increased safety
during the deployment of the autonomous driving system.
It is beneficial to develop, document, and maintain
employee and customer education and training to address
the anticipated differences in the use and operation
of the autonomous driving system from those of the
conventional vehicles that the public owns and operates
today. In the Public Engagement section, we take a deeper
dive into our practices of conducting customer education
and training for safe self-driving testing and future self-
driving deliveries.
Federal, State, and Local Laws
AutoX ensures that our autonomous driving system
adheres to all applicable Federal, State, and local laws. We
will promptly update our system to reflect changes to the
laws. We discuss this in the Federal, State, and Local Laws
section.
OUR SELF-DRIVING TECHNOLOGY 9
18
26
35
39
How Our Self-driving Car “Sees“
The Brain of Our Cars
Fallback (Minimal Risk Condition)
Data Recording
Cybersecurity
10-11
12-15
16
17
17
19
20
21
22
22
23-25
27-28
29-33
34
34
36
36
37
37
38
Sensor Redundancy
Algorithm Redundancy
A.I. Redundancy
System Redundancy
Hardware Redundancy
Human Remote Operator
Operational Design Domain
Testing and Validation
Safety Driver and Remote Operator Training
Federal, State, and Local Laws
Human Machine Interface (HMI)
Showing the Driving Status of the Vehicle
Forced A.I. and Forced Drive-by-wire Disengagement
Post Crash Behavior
Public Engagement
FULL STACK REDUNDANCY
TESTING & VALIDATION
EVERYONE’S SAFETY IS OURPRIORITY
CONCLUSION
CONTENTS INTRODUCTION 3-7
The AutoX Safety Factor 9
Safety has been at the forefront of our minds from the very beginning. We
believe that a safe autonomous driving system can be engineered through
comprehensive design, reliable implementation, and extensive stress testing.
By combining selected sensors with cutting-edge A.I., the systems we build
have a high-resolution perception of their surroundings and are capable of
making safe decisions in all kinds of circumstances. AutoX has strict testing
and vehicle deployment protocols, ensuring that every AutoX self-driving car
on the road is in excellent operating condition and ready to provide safe and
reliable transportation.
Safety in Design
OUR SELF-DRIVING TECHNOLOGY
CHAPTER 1
The AutoX Safety Factor 10
HOW OUR SELF-DRIVING CAR “SEES“
AutoX’s innovative camera perception module provides a
robust, highly accurate object detection and classification
system of the surrounding environment. The combination
of high frame rates, high resolution, and rich color
information makes cameras the ideal backbone for our
sensor suite. Our cameras are critical for mapping, lane
detection, the recognition of traffic lights and traffic signs,
and many other tasks.
The Radar system adds another layer of redundancy onto
the perception module. Radar utilizes electromagnetic
waves to detect the position and speed of objects and
provides us with an effective method of providing better
coverage.
Light Detection and Ranging (LiDAR) is an active sensor
that emits laser beams. A 3D point cloud is generated by
tracing the reflection of laser beams that strike objects. This
is combined with the information provided by the camera
perception system, which creates a layer of redundancy
for our self-driving systems. For instance, pedestrians,
bicyclists, and others on the road are detected with the
combination of LiDAR and camera perception.
Having a well-designed fleet management system and
stable communication with our vehicles is essential as it
allows AutoX to monitor and remote control every one of
our cars. The vehicles will maintain both data and radio
communication links to our fleet command center.
Camera Vision
Radar
LiDAR
Network Antenna and GPS
The AutoX Safety Factor 11
AutoX has designed a human-machine interface that allows other road
users to interact with the vehicle. A driving status display screen at the back
of the vehicle informs the public at all times whether or not the vehicle
is currently in self-driving mode. While our company’s ethos is to ensure
the safety of everyone that our vehicles share the road with, emergencies
and unexpected issues are a possible reality that we must be prepared
for. We have implemented an emergency stop button that enables law
enforcement to bring the vehicle to a complete stop in a safe manner. The
officer or responder may contact the vehicle’s remote operators and take
control of the vehicle if needed.
Driving Status Display and Emergency Stop Button
AutoX puts safety in the hands of the public by allowing them to safely disengage the autonomous mode if necessary.
The AutoX Safety Factor 12
The world’s most talented experts in computer vision, autonomous
driving, and robotics come together at AutoX to develop the brain
of our cars. We have developed numerous innovative methods in
the fields of visual perception, robotics, deep learning, and other
key areas of artificial intelligence. The AutoX team strives to find
ground-breaking solutions to unprecedented problems and share
an unwavering pursuit of technological excellence. The result has
been a unique combination of independent research and safety-
oriented engineering. With our passion and dedication leading
us, we will continue to maintain our rapid progress as we take
driverless vehicles from concept to reality.
The Brain of Our CarsWe are building the brain of a reliable autonomous vehicle to serve our community safely
Inside the Self-Driving Car’s Brain
The localization module pinpoints the vehicle’s exact location. AutoX creates its in-house HD 3D mapping and localization technology and uses real-time sensing to ensure that the maps are up to date.
The perception module allows the vehicle to be informed about its dynamic surroundings. Accurate object recognition and scene understanding are some of the keys to achieving safety. AutoX uses multi-sensor fusion to accomplish this.
The prediction module anticipates other road users’ behaviors and trajectories. This information is then provided to the decision and planning module.
AutoX’s self-driving car is analyzing and optimizing every second to make the safest decision.
The control module takes consideration of the kinematics and dynamics of the vehicle and calculates how much brake, throttle, and steering should be applied.
Localization
Prediction
Perception
Decision & Planning
Control
The AutoX Safety Factor 13
Perception is one of the most challenging tasks for a self-driving vehicle. Our roads are shared by human drivers, cyclists, pedestrians and others, so it is imperative that autonomous vehicles be able to identify, characterize, and react appropriately to them. Road users and other objects are frequently located close to or in front/behind one another - a typical situation is a person standing next to a tree by the side of the road. The eyes of a human driver can easily identify that the pedestrian and the tree are separate objects, and our minds can prepare for the possibility of the pedestrian crossing the road. One major challenge for the autonomous vehicle is to distinguish and recognize each individual object. We leverage our extensive research and development in multi-sensor fusion technology, which provides high precision perception for the A.I. to “see“ reliably and accurately.
Our high resolution sensing goes beyond merely identifying objects. The perception module provides detailed information regarding an object’s orientation, speed, and motion trajectory. This is invaluable information for the prediction module to forecast future motion and informs the next move of the vehicle accordingly. This ability to understand our current environment and predict its state in the immediate future using our perception module’s high resolution sensing greatly enhances the safety of our vehicles.
We have Created a High Resolution Perception Module
At AutoX, we use a variety of carefully selected sensors
to provide us with a holisitic and complete picture of our
vehicles’ surroundings. We combine the high resolution
and rich color information of cameras with the ability
of LiDAR to accurately measure distance and speed.
Coupled with other sensors such as radar, we have
identified a complementary set of sensors, effectively
giving our autonomous cars multiple “senses”.
Different sensors provide layers of redundancy to one
another. We frequently cross-check their readings to
ensure that no object is left undetected and to eliminate
false positives. We are also able to lean more heavily on
just one of them when the other is rendered less effective,
allowing our system to be robust to changing conditions.
Dual main sensors for redundancy
Using our perception module with multi-sensor fusion technology, our A.I. is able to distinguish the flagger and the dynamic traffic sign he is holding. Our system can also detect if the sign is flipped to read “SLOW” insead of “STOP”.
The AutoX Safety Factor 14
Camera image segmentation is the process of analyzing
and detecting different objects from camera inputs and
segmenting the image by finding out what object each pixel
of the image belongs to, such as trees, sky, different types of
vehicles, roads, and road markings. AutoX’s system is able to
analyze and segment camera inputs in real time.
3D object information is critical in the perception
module of autonomous vehicles. We use 3D information
to calculate the dimensions, orientations, and velocities
of surrounding objects. The AutoX approach leverages
color information provided by cameras, 3D information
from LiDAR, and information from other sensors to
provide the decision-engine with enough data to make
safe decisions.
Realtime object recognition
Where? Which way? How fast?
Testing of our real-time segmentation algorithm in highly occluded scenarios in downtown San Francisco.
Segmentation results: the driveable area is shown in pink and the lane detection is shown in cyan. The lane marking detection result is stable under direct sunlight and other lighting conditions.
Based on the camera image, we are able to recognize the heading of the car which helps us to obtain more accurate orientation and velocity information of other cars. The rear of each vehicle is encased by a red square and the front in a green square.
The AutoX Safety Factor 15
We have created systematic methods for multi-sensor
calibration and synchronization. By synchronizing data from
multiple sensors such as cameras and LiDARs, our self-
driving cars are able to see the world in 3D point clouds with
rich semantic information.
HD 3D maps provide detailed road information for the
autonomous vehicle. AutoX combines the data from
multiple sensors including cameras, LiDARs, GPS, and
IMU (Inertial Measurement Unit) to build large scale HD
3D maps with multi-channel information. The larger the
map is, the more challenging it is for the map construction
to be fully autonomous. For example, when the same
intersection is scanned twice by the mapping vehicle, the
algorithm must recognize that the two scan results are
for the same intersection; the map will otherwise contain
an error. We’ve been tirelessly working to solve such
technical challenges. Over the past two years, we’ve built
an automated loop closure system that is more robust
and efficient for large scale HD map building.
Sensor Fusion Large-scale High Definition 3D Maps
Camera + LiDAR fused image
AutoX team built the HD 3D map for North San Jose near our home
The AutoX Safety Factor 16
AutoX’s autonomous driving system has a fallback strategy
that is triggered if the testing vehicle is operating outside
the operational design domain (ODD). A series of reacting
plans will be executed when such situations are detected.
AutoX’s system features a reliable out-of-ODD detection
and is designed to ensure that the minimal risk fallback is
always enabled.
FALLBACK (MINIMAL RISK CONDITION)
There are two aspects of the out-of-ODD detection
in AutoX’s system. First, our team has developed
a comprehensive vehicle health monitor that runs
alongside the A.I. software. The health monitor is
able to detect errors in the message flow amongst
the various software and hardware modules of the
autonomous vehicle system. If errors are detected, the
health monitor will notify the A.I. system and provide
an audible notification to the safety driver or a remote
takeover request to the remote operator. Second, every
AutoX testing vehicle is constantly alerted of the change
of environment by the A.I., a safety driver, or a trained
remote control operator. If any environmental condition
changes are detected (e.g. a sudden heavy rain), the
safety driver and/or remote operator will be able to take
over the testing vehicle or activate a pre-programmed
automatic minimal risk condition fallback.
AutoX’s minimal risk condition fallback strategy
consists of three parts: autonomously attempting to
pull the vehicle over, notifying the safety driver and/
or remote operator, which allows vehicle takeover.
When an out-of-ODD situation is detected or an
out-of-ODD status is received by a remote operator,
the minimal risk condition fallback process will be
activated, commencing the process of autonomously
pulling the vehicle over. This relies on minimal sensors
and software components to come to a safe stop, and
should therefore be operational even if some part of
the A.I. system is not functional. AutoX’s system will
then notify the safety driver or the remote operator
immediately when beginning the process of pulling
over. The safety driver or the remote operator can
take over the testing vehicle at any time.
Out-of-Operational-Design-Domain Detection
Minimal Risk Condition Fallback
The AutoX Safety Factor 17
AutoX has developed a powerful black box system for data
recording during autonomous vehicle testing. The black
box system is able to generate detailed datasets of a whole
test drive, or span three minutes before and after triggering
events. The black box system can be triggered manually by
a safety driver or automatically by the A.I. system. The data
recorded by the black box system includes vehicle position,
velocity, raw sensor input from the LiDAR, cameras, radars,
and GPS along with the corresponding perception outcome,
vehicle decision, planning, control data, and other vehicle
logged data. AutoX uses this information to reconstruct
test scenarios in our simulator, so our self-driving cars are
constantly learning from every mile of driving data collected
by our fleet vehicles. In the event of collisions or accidents,
AutoX will keep the corresponding driving data for crash
reconstruction.
AutoX implements security solutions and infrastructure
based on the cybersecurity principles published by NHTSA.
Our vehicles are equipped with a load balancing engine
that uses Quality of Service (QoS) to control bandwidth
usage, providing reliable and stable traffic to communicate
over bonded network links. We share real-time telematics
with the vehicle during all autonomous testing. The data
communications between each vehicle and the command
center are carried out via secure Virtual Private Network
(VPN) tunnels and only our command center can decrypt
and reassemble packets back into their original order. To
further improve our security practices, we have established
a fleet management program to discover new incidents
and vulnerabilities by monitoring and identifying usage and
behavior, thereby ensuring that the transmission of data is
secure and fully controlled.
DATA RECORDING
CYBERSECURITY
The AutoX Safety Factor 18
To ensure safety, it is essential to think a step ahead and always have a backup
plan. AutoX has designed a multi-layer architecture for redundancy in both
software and hardware. We have also introduced the “highest safety redundancy”
- human remote operators - into the self-driving car testing loop. We believe that
by designing a highly redundant system, we can minimize the possibility of failures
occurring.
Redundancy Is Safety
FULL STACK REDUNDANCY
CHAPTER 2
The AutoX Safety Factor 19
SENSOR REDUNDANCY
The world that self-driving cars navigate in is complex and dynamic; for instance, smaller vehicles are
frequently hidden behind large trucks, and it is difficult to distinguish between joggers and bushes by
the roadside. It is therefore critical to design a perception module with multiple sensors and perception
methods. AutoX has developed a customized sensor fusion technology that includes automatic multi-
sensor calibration and hardware-level synchronization. AutoX’s sensor fusion technology is able to
combine the sensing capabilities of our multiple sensors, resulting in reliable and robust perception
with a broad sensing scope.
Perceiving the world with multiple senses
We are able to precisely filter out the LiDAR points corresponding to pedestrians based on the camera segmentation image, which allows us to build a highly safe autonomous driving technology.
The AutoX Safety Factor 20
ALGORITHM REDUNDANCY
By designing a sophisticated planning and decision algorithm
architecture, we ensure that the brain inside the AutoX self-
driving cars always has a backup plan in mind, creating an
algorithm-level redundancy for safety. There are two aspects
to this. First, given a specific planning task, the decision and
planning modules will check various planning inputs based on
multiple criteria simultaneously. The self-driving software will
always choose the safest solution that has minimal risks. For
example, the planning software will constantly monitor all traffic
around the testing vehicle, even if it is in a protected left turn,
minimizing the risks caused by red light runners. Secondly, no
matter what the current task is, the planning and decision module
will always calculate a safe pullover route, rendering it ever-ready
to respond to an emergency.
We always have a backup plan
The AutoX Safety Factor 21
A.I. REDUNDANCY
At AutoX, we believe that achieving safe and reliable
Level 4 self-driving requires our system to fully utilize
the information provided by our HD 3D maps while
constantly sensing the environment in real time to
correct mistakes and out-of-date maps. When the
results from these two A.I. approaches are properly
fused together, they improve and correct each other’s
errors. This dual A.I. design also ensures that AutoX’s
vehicles will not rely exclusively on one type of sensor or
self-driving method, creating an A.I.-level of redundancy.
Running two A.I. systems simultaneously
The HD 3D Map-Based ApproachHD 3D maps are high definition pre-mappings of the driving environment. They provide important details about static road signatures, such as the location of stop signs, trees, and lane markings. During self-driving, this pre-collected data will be compared to the current sensor input and the self-driving car can precisely discern its own location. At the same time, all of the detailed semantic information from the HD 3D maps are utilized by the A.I. However, it is difficult to guarantee the accuracy of the HD 3D map on a large scale. Thus, we need the real-time sensing approach.
The Real-time Sensing ApproachThe high-resolution real-time sensing approach is more similar to a human driving. In the same way we use our senses to locate lane markings and other cars, this approach uses fused data from multiple sensors and feeds it into the perception algorithm, which is running lane detection and other perception tasks in real time. This approach has a high technological barrier because it requires cutting-edge computer vision and A.I., however it provides AutoX with accurate and up-to-date information, helping the self-driving car respond to sudden unexpected situations. Furthermore, the approach is more robust to transient events such as road construction, which is out of the realm of the map-based approach.
A Tale of Two Approaches:
HD 3D Map-Based and Real-time Sensing
The AutoX Safety Factor 22
SYSTEM REDUNDANCY HARDWARE REDUNDANCY
AutoX vehicles feature two main computers. Should one
computer experience a failure, the second computer will
immediately take over the vehicle, pull over, and make
a safe stop. This decentralized dual computing system
utilizes two reliable computing platforms with the same
computing capability. While this ensures computational
redundancy, we also equip each vehicle with redundant
Drive-by-wire control unit that allows the A.I. to drive the
vehicle. This redundant system approach also accounts
for potential failures in the main system. There is a
fallback system available to take control of the vehicle if
a hardware or software discrepancy is detected.
Decentralized Dual Computing and Drive-by-Wire Control Unit
AutoX has designed a fully redundant hardware
infrastructure in our vehicles to handle hardware issues
such as power failures. It includes redundant backup
power, a redundant emergency stop button, and
redundant connectivity devices.
Double Infrastructure for Safety
The AutoX Safety Factor 23
A remote human operator is the “Highest Safety
Redundancy” for the fully autonmous vehicle. AutoX
has developed a powerful and reliable A.I., though
there still exist extreme situations that are difficult
for any A.I. to handle. These are called “edge cases.”
We have built a remote monitoring and teleoperating
system to enable human remote operators to help
and assist the self-driving car at anytime during
testing and deployment.
The “Highest Safety Redundancy”
Road Construction in Progress
Emergency Vehicles Need to Get By
Traffic Guard on Duty
Bad Weather
Car Accident
Force Detour
Defective Traffic Light
Road Obstruction
HUMAN REMOTE OPERATORS
The AutoX Safety Factor 24
There are two ways a human remote operator could help: “remote takeover” and “remote hybrid support.”
Remote takeover allows a human remote operator to
drive the car remotely as if the remote operator is in
the car. Remote operators will have a steering wheel
and brake and throttle pedals that he or she can use to
control the vehicle, much like conventional driving. We
use a minimum of four cameras to get a surrounding
view of the car. In situations such as a emergency vehicle
passing by, the remote operator is able to take control of
the car and “drive” it to the side of the road and prevent
slowdowns and accidents. There is also a continuous two-
way communication link maintained between the testing
vehicle and the AutoX command center. The human
remote operator can react to law enforcement and other
first responders immediately.
Our self-driving system senses the environment in
many ways. Our vehicles are equipped with sensors
such as LiDARs and radars that perceive the world in a
way that humans cannot. However, in some cases, the
A.I. may still struggle with making decisions since there
are so many challenging situations to handle. With this
in mind, AutoX has designed a remote hybrid support
system, which allows the self-driving car to “ask for
help” with certain decisions that it finds challenging. The
human remote operator can also check the A.I. decision
results and correct or overwrite them when unexpected
errors occur.
Remote Takeover
Remote Hybrid Support
The AutoX Safety Factor 25
AutoX’s remote control system is a part of the remote
teleoperating system. It provides us with a unique layer
of redundancy and the ability to help our A.I. handle
edge cases. At the same time, a great deal hinges on the
human remote operator. AutoX ensures that each of our
operators completes a rigorous training program, and we
provide a remote control system that closely mimics the
process of operating a normal car. The operator’s controls
consist of a steering wheel, accelerator, and brake pedal,
much like a conventional vehicle. The controls are carefully
calibrated to mimic the sense of realistic driving feedback,
and the ergonomics of the entire setup are close to that of
a real car.
Building a reliable remote teleoperating system
We are always watching our vehicles and are ready for human engagement during testing, even if there is no one physically inside the vehicle.
Remote control
Remote hybrid support
Due to the importance of the remote control system, a
great deal of care has been devoted to its development. The
remote control system and the remote hybrid supporting
system have a strict real-time constraint. We cannot tolerate
even a half-second delay or glitches in video streaming or
vehicle control. AutoX engineers have ensured that all the
test vehicles are always connected to AutoX command center
reliably through innovations in both hardware and software.
In terms of hardware, we use multiple high speed modem
bonding and bandwidth aggregation technology. Multiple
channels of high speed internet from different carriers are
used together to share the bandwidth, so that we can get
a much higher data uploading speed compared to other
applications. If one modem fails, the payload can seamlessly
be transferred to other modems.
High resolution video can place a burden upon the network
connection, and network failures are a possibility that we
must contend with. As constant connectivity is a must, we
execute a series of strategies to ensure that the remote
operator never loses contact with the vehicle. For instance, in
the event of poor network connection, we adjust the camera
video to fit the network conditions.
The AutoX Safety Factor 26
By establishing strict testing and validation procedures, AutoX ensures that its
self-driving car tests are conducted safely. We have created a virtual world for
the testing of our self-driving system. Every self-driving feature is first tested
in the powerful simulated environment. Afterwards, well-trained AutoX safety
drivers and remote operators are constantly making sure that our on-road
testing is safe for all road users, the surrounding environment, and ourselves.
AutoX’s Policy and Methods for Ensuring Safe Self-driving Vehicle Testing
TESTING AND VALIDATION
CHAPTER 3
The AutoX Safety Factor 27
OPERATIONAL DESIGN DOMAINAt AutoX, we are developing full-stack Level 4 self-driving technology,
achieving a driving mode-specific performance by an automated driving
task1. It is therefore extremely important to understand the capability of
the self-driving car and standardize the operational design domain.
The operational design domain is a closed set of conditions in which
AutoX driverless vehicles will safely operate only if all conditional
constraints are met. The operational design domain consists of six
elements: the geographic area of testing, road type, speed range,
weather conditions, time of day and types of passengers. AutoX will
not perform Level 4 autonomous testing if any of the conditions listed
above are not fulfilled in the testing scenario. The software and hardware
systems that the AutoX driverless vehicles carry on board, as well as
1. SAE International, J3016_201806: Taxonomy and Definitions for Terms Related to Driving Automation Systems for On-Road Motor Vehicles
We will only test our self-driving car within the range of its capability
The AutoX Safety Factor 28
related systems installed in the AutoX fleet management
and control center, are designed to perform dynamic
driving tasks in the predefined operational design
domain. The dynamic driving tasks are characterized by
the fact that, during the driverless vehicle test, the driving
assignments can be assigned and changed on the fly,
including vehicle destination re-routing, vehicle behavior
changing from city road driving to pulling over, and other
driving behaviors. However, all the driving assignments
given to AutoX driverless vehicles are assured to meet the
predefined operational design domain described in this
section.
To achieve absolute safety, if an unexpected situation
occurs during a driverless vehicle test that is outside
the operational design domain, the test will terminate
immediately and a minimal risk condition fallback will
be triggered. A human operator can take control of
the testing vehicle based on different situations. This is
known as a human operator takeover, and can refer to
either remote operator takeover or in-vehicle operator
takeover. The software and hardware systems of the
AutoX driverless vehicles are designed to ensure remote
operator takeover shortly after triggering. The remote
operator will operate the testing vehicle and transition
it to a safe status as soon as possible, resume the test if
the failing conditions are recovered, or wait for in-vehicle
human takeover or law enforcement takeover depending
on the circumstance. Note that the event that triggers
human operator takeover may not necessarily indicate a
failure in the testing vehicle’s A.I. system. For example, if
the remote operator notices that there is an accident in
the adjacent lane, he or she may make take control of the
vehicle as a precaution.
* Map for illustration purpose only
The AutoX Safety Factor 29
TESTING AND VALIDATION
START FROM THE VIRTUAL WORLD
Whenever new features are added to the AutoX full stack
self-driving system, we conduct tens of thousands of tests
to make sure that the new features meet all our safety
requirements. The first step of testing and validation takes
place in a virtual world: the AutoX 3D simulator.
The AutoX 3D simulator is a multifunctional simulation
and verification platform for autonomous vehicles. It is
able to simulate various scenarios, each of which can be
tailored to the particular circumstances that the team wish
to test. The simulator can accomplish five major tasks:
infrastructure building, sensor and perception simulation,
motion planning and control simulation, traffic and
decision simulation, and full stack verification.
Scenario Editing and Unit Tests
Simulated vehicles interact with virtual
environments called scenarios. The quality
of the virtual environments affects not only
the visual appearance of the simulation, but
also the quality of the sensor simulation and
performance of the simulated vehicles.
The AutoX 3D simulator scenarios are diverse,
with a variety of road types, buildings, traffic
signs, and landscapes. The scenarios are
designed to have a certain level of complexity
and realism for the sensor simulation and
localization simulation, which gives an intuitive
visual feedback to our team as well, helping
them to gauge the performance of the A.I.
algorithms in the virtual world.
3D scenarios can be easily transformed to
HD 3D maps that can be used directly in the
A.I. system, and vice versa. It can also provide
multi-sensor perception results which can be
used for control and motion planning. AutoX
engineers have tuned parameters carefully
to obtain a realistic vehicle dynamic model
that simulates engine RPM and torque output
based on different throttle and brake inputs.
Different road conditions and tire friction
conditions are simulated as well, providing
a wide range of realistic unit test scenarios
for control algorithm and motion planning
algorithm development and refinement.
In the AutoX 3D simulator, scenarios can be created using real HD 3D map collected in the real world or new virtual cities that do not exist by a powerful “world editor.”
The AutoX Safety Factor 30
Creating Scenarios and Edge CasesDecision making is one of the most challenging tasks in the A.I. development of an autonomous
vehicle. There are infinite edge cases that may be difficult or dangerous to reproduce in reality,
such as illegal driving behaviors or sudden traffic accidents. The AutoX 3D simulator provides
a systematic tool to define and produce those traffic scenarios in the virtual world. Other
vehicles and pedestrians are simulated, and their behaviors can be defined or controlled using
descriptive languages and other human-machine interfaces.
The simulated agents in the AutoX 3D simulator can interact with each other and the simulated
self-driving car. They too have “minds“ of their own and are able to follow (or not follow!)
the traffic rules like a real driver, including waiting or running red lights and reacting to stop
signs. Users can ascertain how the self-driving car should behave, for instance when a reckless
driver suddenly cuts in, by reproducing and testing under this situation many times in the
virtual world. In addition to the unit feature development, we can also create more complex
traffic scenarios in the AutoX 3D simulator and have the whole prediction, decision, planning,
and control A.I. in the loop. Algorithms are tested inside the virtual world with many of other
vehicles and pedestrians simulated.
The AutoX Safety Factor 31
Closed Loop Simulation: Adding Sensor Perception to the Simulation
The virtual world is the ideal ground for initial testing and verification of new features.
For this to be effective, however, it is essential that the perception, planning and decision
making modules of the virtual vehicles closely resemble that of the real ones. In the AutoX 3D
simulator, the entire Level 4 self-driving system is run in the virtual world. The 3D simulator
is able to produce simulated camera, LiDAR, and other sensor outputs for the virtual vehicle,
ensuring that our perception system is effectively replicated in the simulation. This way, we
can test the ability of the system to see and classify various objects. This is especially beneficial
in testing the system performance in challenging circumstances, such as situations where a
vehicle is obscured behind a larger one, or pedestrians are standing near trees. Our team
can then optimize the system’s performance and test in these scenarios as many times as is
necessary - all in the safety of the virtual world. By the time a new perception module feature
is ready to be tested in a closed environment in the real world, it has already undergone many
cycles of optimization in the virtual world.
The AutoX Safety Factor 32
Every one of our testing vehicles
goes through extensive pre-flight
safety checks before getting on the
road. These consist of an inspection
performed by our technicians as well
as pre-flight checklists performed by
our autonomous vehicle testers.
Mechanical inspections are performed
on each car to ensure that they are
safe for testing on public roads.
Technicians thoroughly inspect all
safety-related measures such as tire
tread, pressure, headlights, tail lights,
windshield, mirrors, etc. Cars that
do not pass this inspection are made
unavailable for testing until the issues
are resolved.
Pre-flight checklists are completed
every time before the car leaves the
garage for a test on public roads. We
standardize the procedure to ensure
all layers of components associated
with getting an autonomous vehicle
online are tested in the accurate
sequence to ensure the holistic
system is in proper conditions. The
entire procedure includes sensor suit
function checks, loading of the proper
software version, localization checks,
object detection checks, etc.
Pre-flight Safety Checks BEFORE WE GO: VEHICLE MAINTENANCE AND PREFLIGHT CHECK
We constantly monitor the components and the overall
system of each vehicle. We analyze and test our system
at multiple layers, including inspections on the vehicle
itself: vehicle serviced for tire rotation, alignment,
oil & air filter changes, brakes, and other car related
inspections, as well as sensor suit checking, computers
and Drive-by-wire (DBW) hardware checking. Some of
these services are carried out by external professionals
and are strictly monitored by our technicians.
Vehicle Maintenance
Mechanical inspections
Pre-flight checklists
The AutoX Safety Factor 33
FROM CLOSED TESTING SITES TO PUBLIC ROADS
After extensive testing and validation in simulation, AutoX engineers will rigorously
test the integration of the system into a test vehicle. We initially conduct unit tests
in closed testing facilities. Well-trained AutoX employees will create a series of
different scenarios mimicking real driving situations. Various regression tests from
speed control and steering control to the performance of the A.I. components will
be arranged and carried out in a systematic fashion. With these tests, we can
ensure that the vehicle operates safely within our ODD. Only then will we conduct
public road tests. A strict safety protocol will be followed for both manned and
unmanned testing of our self-driving cars. All of the self-driving testing vehicles
will be constantly monitored by a well-trained safety driver and a human remote
operator.
Data-driven Development Approach
The AutoX Safety Factor 34
Safety drivers and remote operators are the last line of
defense in autonomous vehicle safety. AutoX requires
its safety drivers and remote operators to complete
a training program tailored to ensuring proficiency in
identifying problems and controlling the vehicle in a
safe manner. This formal process takes 3 weeks for
each remote operator and 2 weeks for California DMV-
certified autonomous vehicle testers to complete.
Remote operator candidates must also complete this
process, along with additional training specific to remote
teleoperation tasks.
AutoX safety drivers and remote operators understand
that safety is our number one priority, and will follow our
safety policy at all times while operating our autonomous
vehicles. The operator is required to be attentive and
take control of the vehicle in the event of an emergency
and bring it to safety. Operators shall not text, talk on
the phone, eat, smoke, or do anything that they would
not do when driving an actual vehicle. Their attention
must be focused on the road, and they should be in
position to take control of the vehicle at any moment.
AutoX Assistance Operators shall obey all provisions of
the Vehicle Code and local regulation applicable to the
operation of motor vehicles. Test drivers are trained and
required to record field data and create daily reports.
These reports should reflect events completely and
truthfully.
We have designed a systematic training program for
AutoX safety drivers and remote operators. The program
primarily consists of five parts: understanding the
operational design domain, the autonomous system,
the remote control system, fleet management system
user interface training, and field training. We have also
designed an examination with both written tests and
field tests to make sure that every AutoX safety driver
and remote operator is qualified to ensure the safety of
the public and of our self-driving cars.
SAFETY DRIVER AND REMOTE OPERATOR TRAINING
FEDERAL, STATE, AND LOCAL LAWS
AutoX designs its vehicles to meet all Federal,
State, local laws, and any other government
standards. Our vehicles meet the Federal Motor
Vehicle Safety Standards and other relevant
requirements. We also actively communicate
with local authorities, local law enforcement
agencies, and local communities to ensure
that we follow all the government laws and
instructions in terms of our operational and
business practices.
The AutoX Safety Factor 35
AutoX is preparing for driverless vehicle testing. We have devoted a great deal
of effort to building a stable, reliable, and failure proof self-driving software and
hardware system. We are also actively communicating with local law enforcement
agencies and local communities to standardize a series of communication
protocols and implementations. We will specifically address public safety and the
safety of the law enforcement during their engagement with our self-driving cars.
Designing a standardized protocol for public safety
EVERYONE’S SAFETY IS OUR PRIORITY
CHAPTER 4
The AutoX Safety Factor 36
HUMAN-MACHINE INTERFACE (HMI)
AutoX’s self-driving cars will become part of the communities that
they serve, and their presence on the roads calls for a means for other
road users to communicate with them. Our approach to this is three-
fold:
1 We clearly communicate the status of the vehicle at all times, via
a screen at the back of the vehicle. This allows bystanders to be
aware of the state of the self-driving system.
2 We enable law enforcement and other first responders to bring
the vehicle to a safe stop and disengage from autonomous
driving mode in the event of an emergency. This is accomplished by
using either of the emergency vehicles inside or outside the vehicle.
3 We provide an on-board radio as well a cellular phone to
enable contact with vehicle’s remote operator, who can offer
information and support to first responders. We can also operate the
vehicle remotely as required by the situation.
SHOWING THE DRIVING STATUS OF THE VEHICLE
There are three possible driving statuses of AutoX self-driving vehicles: Auto-drive mode, Remote Control mode, and Manual Drive mode.
In Auto-drive mode, the vehicle is controlled by the self-driving software. This indicates that the system is healthy; there are mechanisms in place to detect any anomalies in the self-driving system that can trigger a remote takeover. This mode indicates that AutoX’s level 4 autonomous driving technology is in full control.
In Remote Control mode, the AutoX self-driving car is remotely controlled by a human remote operator. The remote operator is able to bypass the self-driving software and control the vehicle from the command center. The remote control mode is designed as a fallback mode for edge case handling and emergency takeover.
In Manual Drive mode, the testing vehicle is fully controlled by a human driver inside the vehicle. The autonomous vehicle is effectively a normal motor vehicle; no self-driving is taking place.
We have designed a driving status display located on the rear side of the vehicle, helping first responders and other road users to identify the driving status of the vehicle.
Exterior emergency button and A.I. control disengagement switch
The AutoX Safety Factor 37
FORCED A.I. DISENGAGEMENT AND FORCED DRIVE-BY-WIRE DISENGAGEMENT
POST-CRASH BEHAVIOR
In the event of a collision or accident, the self-driving
system may no longer function as expected due to
hardware damage. AutoX has designed an emergency
Human Machine Interface (HMI) to protect law
enforcement officers and other first responders who
may engage with the autonomous vehicle. These consist
of interior and exterior emergency stop buttons and
emergency drive-by-wire disengagement switches. When
conducting a drive-by-wire disengagement, the testing
vehicle will no longer be controlled by any self-driving
software, including the remote control system. This way,
the law enforcement officers and other first responders
can safely interact with the self-driving car, and move it
out of an accident scene to prevent any further collisions.
AutoX has designed an emergency protocol defining
the post-crash behavior of our self-driving testing
vehicles. We are also publishing an AutoX driverless
vehicle law enforcement engagement plan for the
reference of law enforcement agencies and the general
public. AutoX believes that a clear, transparent, and
standardized interaction protocol for any emergency
scenarios can prevent secondary accidents caused by
miscommunication and misunderstandings.
All AutoX driverless test vehicles will be constantly being
monitored by an AutoX remote control operator during
the self-driving car testing. If an AutoX self-driving test
vehicle is in an accident or if a collision is imminent,
the Absolute Emergency Braking (AEB) system will be
activated and the vehicle will come to a full stop. The
remote operator will assess the collision. If there is a
need to move the vehicle to the side of the road, and the
vehicle is in a condition to do so, the remote operator
will do so via remote control. If there is an accident
involving other road users or if there is damage to public
properties, the remote operator will call 911 to report the
accident. The AutoX fleet management team will arrive at
the collision scene as soon as possible to provide on-site
support for law enforcement.
It is of upmost importance to us that our vehicles be
safe, courteous users of the roads. We will closely
cooperate with law enforcement as we continuously test
and perfect our technology.
The AutoX Safety Factor 38
AutoX prioritizes above all the safety of everyone we share the road with. In order to achieve
this, we communicate transparently with our users about our technology and how we handle
emergencies. In the AutoX app, we will provide a detailed description of the vehicle systems
and how the customer should approach and interact with them. The vehicle itself provides
addition instruction through clear, concise instructions posted near its emergency button.
This safety report is meant to serve as an introduction to the operation of our vehicle and
provide a deeper dive into our safety systems. AutoX will continue to produce such content to
share with the public.
All AutoX operators are ambassadors and representatives of our vehicles. They are well-
trained and up-to-date with the operation protocol and our technology. They are ready to
interact with the public to address any questions about our autonomous vehicles.
PUBLIC ENGAGEMENT
The AutoX Safety Factor 39
AutoX has a vision to connect people using our autonomous vehicle technology. The ability
to autonomously transport physical objects opens a whole new world of possibilities.
It promises to save more time that will no longer be lost to traffic and errands. As we
strive to accomplish this, we are cognizant of the fact that our vehicles will become part
of the communities that they serve. As such, the safety of our customers, of other drivers,
pedestrians, and all other road users is the highest priority. This calls for a system that
has safety and reliability engineered into its very fabric - a product that is engineered
from the very beginning with safety in mind. In this report, we have sketched out the
myriad of strategies we have undertaken to ensure the safety our vehicles, from having
high resolution sensing systems to redundancy in both software and hardware to human
oversight via our remote teleoperating system. We have made significant headway in
creating robust, reliable self-driving cars, and we will strive for perfection in every detail.
The road ahead is full of promise, and safe AutoX autonomous vehicles are the best way to
reach the future.
CONCLUSION